Evaluation of gel-pad oligonucleotide microarray technology by using artificial neural networks

Abstract

Past studies have suggested that thermal dissociation analysis of nucleic acids hybridized to DNA microarrays would improve discrimination among duplex types by scanning through a broad range of stringency conditions. To more fully constrain the utility of this approach using a previously described gel-pad microarray format, artificial neural networks (NNs) were trained to recognize noisy or low-quality data, as might derive from nonspecific fluorescence, poor hybridization, or compromised data collection. The NNs were trained to classify dissociation profiles (melts) into groups based on selected characteristics (e.g., initial signal intensity, area under the curve) using a data set of 21,044 profiles derived from 186 probes hybridized to a study set of RNA extracted from 32 microbes common to the human oral cavity. Three melt profile groups were identified: one group consisted mostly of ideal melt profiles; another group consisted mostly of poor melt profiles; and, the remainder were difficult to classify. Screening of melting profiles of perfect-match hybrids revealed inconsistencies in the form of melting profiles even for identical probes on the same microarray hybridized to same target rRNA. Approximately 18% of perfect-match duplex types were correctly classified as poor. Experimental variability and deviation from ideal melt behavior were shown to be attributable primarily to a method of local background subtraction that was very sensitive to displacement of the grid frames used for image capture (both determined by the image analysis system) and duplexes with low binding constants. Additional results showed that long RNA fragments limit the discriminating power among duplex types.

FIG. 1.

Example criteria used to classify melting profiles. A to E, Actual melting profiles; F, Random profile. Quality criterion: A and C have low scores because they are smooth; B and D to F have disjointed points as indicated by the arrows and have high scores. Shape criterion: A and B have low scores because they have ideal shapes.

FIG. 2.

Importance of NN inputs to predict Quality (top panel) and Shape (bottom panel) scores relative to an idealized melting profile. The relative importance of each input was based on rank order and their corresponding P_model and E ratio values (see text). Shaded areas (and solid black circles) indicate inputs that were statistically more important than the other inputs (white circles). Note that for Shape scores (lower panel), NN inputs 26 and 27 were also found to be important for predicting Shape scores. Inputs 26 and 27 correspond to Quality 1 and 2 scores, respectively.

FIG. 3.

Ordination plots produced by PCA of melting profile variables. P, perfectly matched probe target duplexes; I, duplexes containing an internal mismatch; N, duplexes containing more than two mismatches.

FIG. 4.

Melting profile performance scores by position on the ordination plot. Top panel, two principal components; bottom panel, absolute duplex performance scores (predicted by MPP calculator) relative to PCA1. Blue dots represent melting profiles with duplex performance scores of >0.75; red dots represent profiles with scores between 0.25 and 0.75; black dots represent profiles with scores of <0.25.

FIG. 5.

Interpretation of melting profiles for perfectly matched probe-target duplexes by the MPP calculator. Top panel, a probe (Univ 1390) that tends to yield high initial signal intensity values; lower panel, a probe (S-P-Grpos-1200-a-A-13) that tends to yield low initial SI values. Ideal profiles, blue lines; uncertain profiles, red lines; and poor profiles, black lines. Not all melting profiles are shown for clarity.

FIG. 6.

Distribution of initial SI values of perfectly matched probes from data set 1 by MPP type. Initial SI values of all perfectly matched melting profiles are shown as shaded background (i.e., including uncertain profiles). The range, mean ± standard deviations (gray) of initial signal intensities for probes 438 (S-P-Grpos-1200-a-A-13, ΔG⁰ = −21.8 kcal/mol) and 62 (Univ 1390, ΔG⁰ = −33.8 kcal/mol) are presented in the horizontal bars above.

FIG. 7.

Relationship between initial SI values of perfectly matched duplexes and their corresponding binding constants (ΔG⁰₂₀) for solution. SI values were determined by the Fotin et al. (13) method. Mean and standard deviation (n > 40) of SI values are shown for each duplex. Probes in order from lowest to highest ΔG⁰₂₀ are (number, name): 63, Univ 907; 64, Eub 927; 65, Eub 338; 390, Eub 336; 74, S-P-Grpos-1192-a-A-22; 62, Univ 1390; 75, S-P-Grpos-1199-a-A-15; and 438, S-P-Grpos-1200-a-A-13.

FIG. 8.

Distribution of initial SI values of probes having more than two mismatches to target sequences from data set 1 by MPP type. Initial SI values of melting profiles of all probes with more than two mismatches are shown as shaded background (i.e., including uncertain profiles).

FIG. 9.

Color-enhanced image of a portion of a gel-pad microarray showing the inner and outer grids framing the gel pads used by the Fotin et al. (13) image processing software. Gel-pads, green; frames, pink; background, blue. Panel B is a magnified image of single gel-pad from panel A.

FIG. 10.

Composite melting profiles derived from the analysis of 20 displaced image stacks from a single gel-pad microarray experiment. Placement of the original stack image was displaced in the by and y coordinates of the frame (see text for details). Panels A, C, and E represent stack images processed by using different background subtraction methods: in-out/out (equation 4) (13), the in-out (equation 5) (52), and the in method (equation 6) (this study), respectively. Panels B, D, and F represent the normalized melting profiles (used to calculate the dissociation temperature) from panels A, C, and E, respectively. Gray boxes indicate the range of T_d values.

FIG. 11.

Differences in melting profiles obtained using synthetic (i.e., oligonucleotide) (panels A and C) and native rRNA target (panels B and D). Images were obtained using equation 6. Panels A and B represent probes 62 (Univ 1390, perfect match, PM) and 399 (Univ 1390-c13, single internal mismatch, MM), respectively. Panels C and D represent probes 63 (Univ 907, PM) and 401 (Univ 907-c9, MM), respectively.